一种基于自校准知识蒸馏的低照度图像增强方法
Self-Calibrated Knowledge Distillation for Low-Light Image Enhancement
-
摘要: 低照度图像增强是计算机视觉领域重要研究内容之一. 为了增强低照度图像并显著减少模型的参数量并有效地抑制噪声和增强图像细节, 提出一种基于自校准知识蒸馏的低照度图像增强方法——SCKD. 首先将异构的知识蒸馏与Retinex理论相结合, 提出Retinex-KD框架, 通过规范化教师网络蒸馏条件与学生模型增强步骤, 引导教师模型将亮度、色彩和纹理等细节信息传递给学生模型, 提升增强细节; 然后提出低照度增强学生网络LDFC-Net, 包括照明引导校准(LGC)模块与光照干扰抑制校准(LISC)模块, 其中, LGC模块恢复低照度图像照明并增强细节, 得到光照估计特征图, LISC模块抑制光照估计特征图的噪声, 使得图像更真实; 最后设计专用蒸馏损失函数, 并结合Retinex-KD对LDFC-Net进行蒸馏, 有效地降低学生模型参数量并保证增强性能. 在LOL-v1和LOL-v2-real数据集上的实验结果表明, 与主流轻量化方法相比, SCKD在PSNR和SSIM指标上分别平均提高1.862 dB和4.15%, 且仅需50 K的参数量和2.98 G的计算量就能实现与非轻量化主流方法相接近的性能, 效率大幅提升.Abstract: Low-light image enhancement is an important research area in computer vision. To reduce model parameters significantly while effectively suppressing noise and enhancing image details, this paper proposes a Self-Calibration Knowledge Distillation method for low-light image enhancement——SCKD. First, heterogeneous knowledge distillation is combined with Retinex theory to propose the Retinex-KD framework, which normalizes the teacher network's distillation conditions and the student model's enhancement steps, guiding the teacher model to transfer brightness, color, and texture details to the student model, thereby improving enhancement details. Second, a low-light enhancement student network called LDFC-Net is proposed, which includes a Light-Guided Calibration (LGC) module and a Light Interference Suppression Calibration (LISC) module. The LGC module recovers low-light image illumination and enhances details to obtain an illumination estimation feature map, while the LISC module suppresses noise in the illumination estimation feature map, resulting in more realistic images. Finally, a specialized distillation loss function is designed and combined with Retinex-KD to distill LDFC-Net, effectively reducing the student model's parameter count while ensuring enhancement performance. Experimental results on the LOL-v1 and LOL-v2-real datasets show that, compared to mainstream lightweight methods, SCKD improves PSNR and SSIM metrics by an average of 1.862 dB and 4.15% respectively, while requiring only 50 K parameters and 2.98 G computations to achieve performance comparable to non-lightweight mainstream methods, significantly improving efficiency.